Maybe need to pay attention to The Precision of Float Type... about frnn HOT 3 CLOSED

Sorry for the late reply, no idea why I didn't get the notification by email...

Thanks for the feedback. If I understand it correctly, you mean the grid size should be [20, 20, 20] but the float-point issue makes it to be [20, 21, 20]?

If that is the case, here is my considerations: in the general case, the size of the bounding box would not be the multiple of grid size / search radius. Then we don't have this kind of issues. Even if this happens, I guess both the computational overhead and the memory overhead is marginal.

Maybe I should change the data type of grid_params to be double? Do you have any suggestion?

from frnn.

I think the "round" operation seems to be more reasonable rather than "int" or "floor"

from frnn.

That's a good idea. But if we round down, the search radius can be larger than cell_size * int_ratio, which means we need to search one more layer of grids (e.g. originally we search over a 5x5x5 grid, now with this round-down, we need to search over 7x7x7). According to my experiment results, this would be much slower if the search radius is not the exact multiple of the cell_size.

from frnn.